Overcoming the biggest blocker to AI production | Tech Radar

Overview

News, deals, reviews, guides and more on the newest computing gadgets

Start exploring exclusive deals, expert advice and more

Details

Unlock and manage exclusive Techradar member rewards.

Unlock instant access to exclusive member features.

Get full access to premium articles, exclusive features and a growing list of member rewards.

Legacy security models stall AI agents. We need unified, zero-anonymity identity

When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works.

Autonomous AI agents are already running inside core infrastructure – executing code, applying policies, and managing Dev Ops functions. And the projects keep stalling, because the security models they’re being wired into were built for a world that no longer exists.

Retrofitting non-deterministic actors into those models is costing engineers time they don’t have, and it introduces risk no enterprise can manage.

Many projects have stalled amid concerns about deploying without a robust security foundation – and with good reason. We’ve already seen an agent delete a company’s entire production database, and its backups, in nine seconds. Security teams are being asked to stop scenarios like that with tools built for a world of two actors. The cracks are starting to show. Something has to change, or innovation stalls under the weight of its own controls.

The pressure on production and engineering teams to speed up delivery is very real and pervasive. So they often fall back on old habits like granting agents broad privileges and treating them as any other microservice.

But agents are very different from machines; they are error-prone and non-deterministic, just like humans. Yet, operating at machine speed, 24/7. Agents can delete entire production databases in nine seconds. How many humans do you know who could do that?

AI agents are inside the enterprise – are your security foundations ready for them?

Artificial intelligence agents need access, not secrets

Why AI coding agents keep stalling before production and the governance controls that fix it

And this brings me to the crux of the problem. Identity systems were built for a world with two kinds of actors – humans and machines – but there are now three. Trying to fit agentic AI into outdated systems makes each agent a potential source of compromise, and one that can execute thousands of actions across infrastructure in seconds.

Yet this is what engineers are asked to do; stop catastrophic scenarios with legacy IAM tools that are breaking down. The cracks are starting to show.

Historically, identity fragmentation has plagued engineers working with Kubernetes clusters, cloud platforms, container orchestration, CI/CD pipelines, databases, etc.

Enter agentic AI, and the speed gets turned up to the max. Suddenly, teams are inundated with thousands of activity logs, and they lack the ability to effectively contain the agent before it executes unauthorized changes.

Agentic security doesn't need a whole new definition – you just need to reframe what you already know

Why cybersecurity must evolve for the age of AI agents

Phishing the agent: Why AI guardrails aren’t enough

Trying to enforce strong authentication and short-lived privileges would mean building individual integrations for every tool in the stack. It makes AI hard to scale when each tool uses a different integration protocol.

Rather than focusing on innovating with agentic AI, engineers are forced to stitch together IAM, infrastructure, and secrets by hand — with no consistent identity, no visibility into agent actions, and every team building its own container or VM workflows from scratch.

But creating a new tool to handle a third identity type is the worst reaction the industry could have. It would double the work for engineers, as they would need to rebuild the identity policy from the ground up and introduce greater anonymity. A new identity silo is anonymous to other siloed systems, making it even harder to catch attackers.

This leaves us with the question of how enterprises can control an agent's behavior. The solution isn’t about adding to the tech stack or implementing more tools; it's about changing our identity models to eliminate anonymity entirely.

To remove anonymity from infrastructure, enterprises must give every actor –spanning humans, machines, workloads, and AI agents – first-class identities, cryptographically secured by a hardware root of trust.

Ditch static credentials entirely. Eradicating API keys and passwords eliminates the credential sprawl that causes breaches, as well as the threat of secrets being stolen or handed over to the wrong actors. With identity rooted in real-world factors, attackers cannot impersonate a trusted machine and trick an agent into exfiltrating a database.

But strong authentication in itself is not enough. AI agents, like all other actors, need to adhere to zero-trust principles. This can only happen when siloed systems are replaced by an infrastructure layer in which agents have the exact same identity type as the machines they run on and the humans who authorize them.

Agents should operate with short-lived privileges tied directly to specific actions authorized by a human user. Privilege attached to the action, not the actor. For example, an agent generating code must inherit its mandate from a human owner with matching authority, restricting privileges to only the specific data tables required for that task.

Non-deterministic actors also require a contained, trusted execution environment before touching production infrastructure.

With no default privileges, the blast radius is heavily bounded. But this can only be achieved when a single policy is set by a single system for all identities.

Identity policy can also be introduced as an enforcement layer between the agent and its inference endpoint, so behavior is controlled before instructions are ever executed.

With a unified architecture, identity becomes the control plane for AI

AI agents are unlocking immense opportunities in enterprise environments, especially when deployed in live infrastructure, where they deliver the most value. From managing routine changes to fixing deployments in real-time, the possibilities are endless. Succeeding with AI in such critical business operations requires tight control of behavior.

This article was produced as part of Tech Radar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of Tech Radar Pro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

You must confirm your public display name before commenting

How to secure your business by moving from passwords to passkeys

How to edit an entire week of social media video on your lunch break with Veed's AI video editor

Dali's colorful new subwoofer is designed to make its affordable 5-star stereo speakers sound even bigger and better — or to be part of a great-value, stylish home theater setup

TP-Link reveals 'world's first Wi-Fi 8 lineup' at IFA 2026 with Archer 9 Ultra router and Deco 8 Ultra mesh Wi-Fi system

Today I learned that F-35 pilots get $400,000 helmets with 'X-ray vision' - but the Pentagon almost ditched them

Tech Radar is part of Future US Inc, an international media group and leading digital publisher. Visit our corporate site.

Key Takeaways

News, deals, reviews, guides and more on the newest computing gadgets
Start exploring exclusive deals, expert advice and more
Unlock and manage exclusive Techradar member rewards
Unlock instant access to exclusive member features
Get full access to premium articles, exclusive features and a growing list of member rewards